self-improve-checkpoint-resume: resume openevolve MAP-Elites population across ~6h cycles

.vcsdd/active-feature.txt

@@ -1 +1 @@
-franklin-alwaysact-skill-router
+self-improve-checkpoint-resume

.vcsdd/features/self-improve-checkpoint-resume/evidence/red-phase.md

@@ -0,0 +1,6 @@
+new-feature-tests: FAIL (pre-implementation, confirmed via git stash of run_evolve.sh + lib/checkpoint_resume.py)
+regression-baseline: PASS
+
+=== RED (stashed impl, tests only) ===
+1 collection error (ImportError: cannot import name checkpoint_resume) blocking test_checkpoint_resume.py's 23 tests
+test_checkpoint_resume_wiring.py: 7 failed, 1 passed (PROP-CR-WIRE1/2/2b/3x2/13/LIVE1 all failed as expected pre-wiring)

.vcsdd/features/self-improve-checkpoint-resume/reviews/converge/output/verdict.json

@@ -0,0 +1,96 @@
+{
+  "phase": "converge",
+  "feature": "self-improve-checkpoint-resume",
+  "verdict": "CONVERGED",
+  "iteration": 1,
+  "reviewerToolConstraints": "This convergence session had NO Bash/execute tool available (only Read/Write/Edit/Grep/Glob). I could not run `python3 -m pytest`, `semgrep`, `bash -n`, or `git diff`/`git log -p` myself. Per the task's own instructed fallback, I cross-referenced verification-report.md's '31/31 pass' and 'full suite 18 failed/102 passed/1 skipped' claims against the on-disk `.pytest_cache/v/cache/{lastfailed,nodeids}` artifacts (still present in this worktree) and against a byte-for-byte count of `def test_` occurrences in both actual test files, rather than re-executing pytest.",
+  "dimensions": [
+    {
+      "name": "spec_fidelity",
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": [
+        {"type": "file", "location": "skills/earn/self-improve/lib/checkpoint_resume.py:15-69", "description": "Direct read this session: find_latest_checkpoint's absolute-path normalization (REQ-CR1, line 52 os.path.abspath), filter->exclude->sort-descending->iterate order (REQ-CR2, lines 56-67), REQ-CR10 run-name regex `^run-\\d{8}T\\d{6}Z$` (line 19), integer (not lexicographic) checkpoint comparison via int(match.group(1)) (REQ-CR3, lines 38-41), os.path.isdir guard before listdir (EDGE-CR12, line 29) all match behavioral-spec.md's literal requirement text."},
+        {"type": "file", "location": "skills/earn/self-improve/run_evolve.sh:75-85,103-110", "description": "REQ-CR6's pinned call string `RUNS_DIR=\"$RUNS_DIR\" RUN_ID=\"$RUN_ID\" \"$PY_BIN\" \"$SKILL_DIR/lib/checkpoint_resume.py\"` is present verbatim at line 75, positioned after $RUN_DIR creation (line 52) and before the first `\"$OPENEVOLVE_BIN\"` occurrence (line 87). All three REQ-CR9 pinned log substrings ('checkpoint_resume: resuming from checkpoint', 'checkpoint_resume: no prior checkpoint found', 'checkpoint_resume: resume-check crashed') are present verbatim in their correct branch (lines 79/81/84). The FIND-001 fix idiom `\"${CHECKPOINT_ARGS[@]+\"${CHECKPOINT_ARGS[@]}\"}\"` is present at line 109, exactly as both impl-review verdicts describe -- independently re-confirmed by direct read this session, not merely trusted from the narrative."},
+        {"type": "file", "location": "skills/earn/self-improve/config.yaml:11,129", "description": "behavioral-spec.md's sources section cites `checkpoint_interval: 10` at config.yaml line 11 and `database.db_path: \"runs/db\"` at config.yaml line 129 (EDGE-CR10). Both independently re-read this session and confirmed byte-accurate: line 11 is exactly `checkpoint_interval: 10  # top-level Config field...` and line 129 is exactly `db_path: \"runs/db\"`."}
+      ]
+    },
+    {
+      "name": "edge_case_coverage",
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": [
+        {"type": "file", "location": "skills/earn/self-improve/tests/test_checkpoint_resume.py:1-418", "description": "Counted 23 `def test_` functions in this file this session, matching verification-report.md's PROP-CR1/1b/2/3/4/5(x2)/6/7/8(x2)/9/10(x2)/11/12a/12b/12c(x2)/14(x2) plus the residual REQ-CR4 test -- every EDGE-CR1 through EDGE-CR12 case has a dedicated tmp_path fixture (EDGE-CR11's stray-.DS_Store-only checkpoints/ dir at lines 274-280, EDGE-CR12's zero-byte plain-file-named-like-a-run-dir at lines 293-303, both hand-verified this session)."},
+        {"type": "file", "location": "skills/earn/self-improve/tests/test_checkpoint_resume_wiring.py:1-343", "description": "Counted 8 `def test_` functions in this file this session (wire1/wire2/wire2b/wire3x2/cr13/live1/regression-syntax) = 23+8=31 total, matching every one of the 31 test node IDs actually present in `.pytest_cache/v/cache/nodeids` (grep-confirmed this session: exactly 31 `checkpoint_resume` occurrences, each an exact match to a real `def test_` name in these two files -- no phantom test name in either document)."}
+      ]
+    },
+    {
+      "name": "implementation_correctness",
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": [
+        {"type": "file", "location": "skills/earn/self-improve/run_evolve.sh:75-113", "description": "Full pre-invocation step hand-traced this session against REQ-CR6/7/8/9/12: exit-status capture (line 76) precedes CHECKPOINT_ARGS initialization (line 77); if/elif/else (78-85) prioritizes the crash branch before found/not-found; the fixed array-expansion at line 109 (the FIND-001 fix) is present verbatim, matching what both impl-review verdicts describe as the reviewed state -- this session's direct read of the CURRENT worktree file (HEAD=d4990326, not a stale copy) shows no drift from what iteration-2's verdict (reviewed_commit eba2270a) described, since Phase 5's only subsequent commit (d4990326, 'Phase 5 formal hardening complete') is stated in state.json's own phaseHistory to have added only verification/*.md files, and this session's direct read of checkpoint_resume.py/run_evolve.sh/both test files independently confirms their content is unchanged from what iteration-2 reviewed."},
+        {"type": "test-output", "location": "skills/earn/self-improve/.pytest_cache/v/cache/lastfailed", "description": "Read this session: exactly 18 entries, every one in a pre-existing, unrelated test file (test_adversary_disapprove.py, test_baseline_beat.py, test_denylist_reject.py, test_heldout_regress.py, test_ledger_reader.py, test_realized_gate.py, test_risk_adjusted_fitness.py) -- ZERO entries from test_checkpoint_resume.py or test_checkpoint_resume_wiring.py, corroborating (independently of the narrative) verification-report.md's '31/31 pass, 18 pre-existing unrelated failures' claim. This cache file is a durable, still-present artifact (unlike the specific .pyc bytecode files iteration-2's verdict also cited -- see structural_integrity finding below for that caveat)."}
+      ]
+    },
+    {
+      "name": "structural_integrity",
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": [
+        {"type": "file", "location": "skills/earn/self-improve/{config.yaml,promote_gate.sh,lib/promote.py,lib/promote_gate.py,strategies/pm_backtest_strategy.py,launchd/ai.anicca.self-improve-evolve.plist}", "description": "Fresh Grep this session for the literal string 'checkpoint_resume' across all six INV-CR2-protected files: ZERO matches in any of them -- independently confirms INV-CR2 ('this feature never touches' these files) holds against the actually-current worktree state, not merely against a prior session's grep."},
+        {"type": "file", "location": "skills/earn/self-improve/lib/checkpoint_resume.py:15-17", "description": "Direct read confirms the module's ONLY imports are `import os`, `import re`, `import sys` -- no subprocess/socket/requests/urllib/shutil/pathlib -- matching PROP-CR9/PROP-CR12a/PROP-CR12b's claims independently, not merely re-stating purity-audit.md's own self-report."}
+      ]
+    },
+    {
+      "name": "verification_readiness",
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": [
+        {"type": "file", "location": "skills/earn/self-improve/tests/test_checkpoint_resume_wiring.py:166-186", "description": "Direct read confirms FIND-002's fix is real and present: `test_prop_cr_wire2_no_checkpoint_found_argument_list_unchanged_and_logged` asserts `len(capture_lines) == 8` and `\"\" not in capture_lines`, not merely substring absence of '--checkpoint' -- genuinely capable of catching a regression of FIND-001's empty-array-expansion bug, independently re-verified this session against the actual test source, not the verdict's own description of it."},
+        {"type": "file", "location": ".vcsdd/features/self-improve-checkpoint-resume/verification/verification-architecture.md:60-79 vs verification-report.md:8-29", "description": "Every one of the 19 PROP-CR* rows in verification-architecture.md's Proof Obligations table has a corresponding 'PROVED' row in verification-report.md citing a real, exact test function name -- cross-checked this session against the actual test file contents (both files read in full) and against `.pytest_cache/v/cache/nodeids`: every cited test name exists verbatim in the real test files and was collected by a real pytest run. No PROP references a test name that does not exist."}
+      ]
+    }
+  ],
+  "criteria": {
+    "finding_diminishment": {
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": "Phase 3 impl-review: 2 blocking findings (iteration 1) -> 0 blocking findings (iteration 2), both independently re-confirmed resolved this session by direct code read (run_evolve.sh:109's fixed array idiom; test_checkpoint_resume_wiring.py:166-186's strengthened assertion). This is this feature's first Phase 6 convergence pass; there is no prior converge iteration to diminish against, but the Phase 3 trend itself (2->0) is genuine, not a renamed/deferred finding."
+    },
+    "finding_specificity": {
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": "Both Phase 3 findings (FIND-001, FIND-002) carried exact file:line evidence and were independently re-derivable this session from the actual current file contents, not merely trusted from the fix narrative -- FIND-001's fix (run_evolve.sh:109) and FIND-002's fix (test_checkpoint_resume_wiring.py:166-186) were both located and read directly, not located via the verdict's own line-number claim alone (I independently re-read the surrounding lines to confirm the described content, not just that *a* line existed at that number)."
+    },
+    "criteria_coverage": {
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": "No contracts/sprint-N.md exists for this feature (Glob of .vcsdd/features/self-improve-checkpoint-resume/{contracts,reviews/contracts}/** returns nothing) -- this feature used lean-mode spec-review + impl-review only, no strict CRIT-XXX contract phase. All 19 PROP-CR* proof obligations in verification-architecture.md are accounted for in verification-report.md's discharge table (cross-checked this session, see verification_readiness dimension)."
+    },
+    "four_dimensional_convergence": {
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": "SPEC (behavioral-spec.md REQ-CR1-12/INV-CR1-5/EDGE-CR1-12, verification-architecture.md's 19 PROP-CR* rows), TEST (31/31 passing per .pytest_cache/nodeids cross-check against real test file contents), IMPL (lib/checkpoint_resume.py + run_evolve.sh's new step, directly re-read this session and matching every cited requirement), and VERIFICATION (verification-report.md/security-report.md/purity-audit.md, whose specific factual claims -- config.yaml:11/129 citations, import list, INV-CR2 file-touch scan, PROP discharge table -- were independently re-derived this session rather than rubber-stamped) are mutually consistent for this feature's entire surface."
+    },
+    "duplicate_detection": {
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": "FIND-001 and FIND-002 (Phase 3 iteration 1) are genuinely resolved, not renamed or deferred: the fix at run_evolve.sh:109 is the exact idiom both impl-review iteration-2 and this session independently identify as the canonical portable fix, and the strengthened test at test_checkpoint_resume_wiring.py:166-186 is demonstrably capable (by its own asserted argument-count/empty-string checks) of catching a regression of the exact bug class FIND-001 was. No new instance of either bug class was found this session in either file."
+    },
+    "residuals_honest": {
+      "verdict": "PASS",
+      "findings": [],
+      "evidence": "Two non-blocking residuals are disclosed here rather than swept into a blanket 'zero deviations' claim: (1) the specific `.pyc` bytecode artifacts iteration-2's impl-review verdict cited as 'independently-discovered... could not have been fabricated' forensic evidence (lib/__pycache__/checkpoint_resume.cpython-314.pyc, tests/__pycache__/test_checkpoint_resume_wiring.cpython-314-pytest-9.1.1.pyc) do NOT exist in this worktree as of this session (confirmed via Glob of skills/earn/self-improve/**/__pycache__/** -- zero matches); this is most plausibly explained by __pycache__ being gitignored (skills/earn/self-improve/.gitignore:3-4) and transient/cleaned since that review session, and the claim is NOT falsified by this absence because the more durable `.pytest_cache/v/cache/{lastfailed,nodeids}` artifacts DO still exist and DO independently corroborate the same underlying '31/31 pass, 18 unrelated pre-existing failures' claim (cross-checked this session against the real test file contents, not merely re-read as narrative) -- so the underlying factual claim survives even though one specific corroborating artifact it cited no longer does. (2) Phase 1c's own spec-review iteration-5 verdict (reviews/spec/iteration-5/output/verdict.json) lists only 2 of the usual 5 review dimensions (spec_fidelity, verification_readiness) rather than a full 5-dimension re-evaluation -- a process-completeness observation about an already-gated (PASS, 0 blocking) prior phase, outside this session's primary Phase 3/5-cross-check mandate, noted here rather than omitted. Neither residual is blocking: no code-level claim in verification-report.md/security-report.md/purity-audit.md that this session checked (config.yaml line citations, import list, INV-CR2 grep scan, PROP-to-test-name mapping, argument-count assertions) was found false."
+    }
+  },
+  "blocking_count": 0,
+  "reviewed_commit": "d4990326 (current worktree HEAD; Phase 3's own reviewed_commit was eba2270a -- the two intervening commits, per state.json's phaseHistory, added only verification/*.md hardening docs, and this session's direct read of the actual current lib/checkpoint_resume.py, run_evolve.sh, and both test files confirms their content is unchanged from what iteration-2 reviewed)",
+  "convergenceSignals": {
+    "findingCount": 0,
+    "allCriteriaEvaluated": true,
+    "evaluatedCriteria": [],
+    "duplicateFindings": []
+  },
+  "ts": "2026-07-12T00:00:00.000Z"
+}

.vcsdd/features/self-improve-checkpoint-resume/reviews/impl/iteration-1/output/findings/FIND-001.json

@@ -0,0 +1,14 @@
+{
+  "findingId": "FIND-001",
+  "dimension": "implementation_correctness",
+  "category": "requirement_mismatch",
+  "severity": "critical",
+  "blocking": true,
+  "description": "run_evolve.sh:109 appends the new checkpoint flag via `\"${CHECKPOINT_ARGS[@]:-}\"`. When no checkpoint is found (or the resume-check crashes), CHECKPOINT_ARGS stays the empty array declared at run_evolve.sh:77 (`CHECKPOINT_ARGS=()`). Under `set -u` (run_evolve.sh:26) on bash < 4.4 -- which is exactly the shell this script's own header comments (run_evolve.sh:17-25) say it targets, and which is macOS's stock, never-upgraded system `/bin/bash` (3.2, unchanged for licensing reasons) that a launchd-invoked script with a minimal PATH will resolve to -- expanding `${arr[@]}` on a declared-but-empty array is treated as 'unset' for nounset purposes (fixed only in bash 4.4's changelog item 'Expanding an empty array is no longer an error when nounset is enabled'). The `:-` operator's fallback, when triggered, substitutes its (empty) word literally as ONE quoted argument -- it does not re-apply `[@]` array-splitting to the fallback the way `${arr[@]+\"${arr[@]}\"}` (the canonical, fully-correct fix for this exact bash pitfall) does. The practical effect: in the 'no checkpoint found' branch (and the 'resume-check crashed' branch), `\"$OPENEVOLVE_BIN\" ... --output \"$RUN_DIR\" \"${CHECKPOINT_ARGS[@]:-}\"` likely passes ONE spurious trailing empty-string argument to openevolve-run, violating REQ-CR8/INV-CR4's explicit 'invoke openevolve-run with EXACTLY the same argument list it uses today ... byte-for-byte the pre-existing behavior'. Confirmed via direct read of the installed CLI's own arg parser (~/.anicca-venvs/self-improve/lib/python3.14/site-packages/openevolve/cli.py:22-26): openevolve-run's argparse defines EXACTLY two positional arguments (initial_program, evaluation_file) and no catch-all/nargs='*' positional -- a third stray positional token (even an empty string) is unrecognized and argparse raises 'error: unrecognized arguments' and calls sys.exit(2). Since the not-found branch is the MOST COMMON path (every first-ever run, and every run until a valid checkpoint accumulates -- which, if this bug fires, would never happen because every such cycle would crash openevolve-run immediately), this is a regression strictly worse than the tied-combined_score stagnation this feature was built to fix: instead of a wasted-but-completing cycle, every cycle in the fallback branch would fail outright. I could not execute a live `bash -c` repro or check this Mac's actual `/bin/bash --version` to settle this with 100% certainty -- this review session has no Bash/execute tool available, only Read/Grep/Glob/Write/Edit -- so this finding is reasoned from (a) documented bash nounset+empty-array semantics and the well-known `:-` vs `+` fallback distinction, (b) the directly-read openevolve argparse source, and (c) FIND-002's confirmation that no existing test actually checks full argument-list identity in this branch. Recommended fix: replace `\"${CHECKPOINT_ARGS[@]:-}\"` with `\"${CHECKPOINT_ARGS[@]+\"${CHECKPOINT_ARGS[@]}\"}\"`, or branch explicitly on `${#CHECKPOINT_ARGS[@]}` before invoking `$OPENEVOLVE_BIN`, then re-verify by actually running the composed command under this machine's real `/bin/bash` and inspecting the captured argv byte-for-byte.",
+  "evidence": {
+    "filePath": "skills/earn/self-improve/run_evolve.sh",
+    "lineRange": "77,103-110",
+    "snippet": "CHECKPOINT_ARGS=()\n...\n\"$OPENEVOLVE_BIN\" \\\n  \"$SKILL_DIR/strategies/pm_backtest_strategy.py\" \\\n  \"$SKILL_DIR/evaluator.py\" \\\n  --config \"$SKILL_DIR/config.yaml\" \\\n  --iterations \"$ITERATIONS\" \\\n  --output \"$RUN_DIR\" \\\n  \"${CHECKPOINT_ARGS[@]:-}\" \\\n  >> \"$LOG\" 2>&1"
+  },
+  "routeToPhase": "2b"
+}